Brain-like computing is a new computing technology based on the development of neuromorphic engineering, drawing on the basic principles of brain science, artificial-oriented general intelligence.
Why develop such a technology? Now human beings live in a digital universe, where everything is connected anytime, anywhere, forming a digital universe where everything is interconnected. This universe is growing very fast, information is doubling every two years, the entire universe is expanding rapidly, and it never regresses, and such a universe is based on our current computer architecture, which in turn is based on the von Neumann architecture.
The Von Neumann architecture is in my opinion the most concise, beautiful and most influential architecture in the history of human development. It is characterized by the separation of computing and storage, and the computing and storage are scheduled back and forth through the bus. It can be imagined that the round-trip scheduling consumes a lot of energy, delays time, slows down, and causes congestion, so there is a bandwidth bottleneck.
2017 Turing Award winner John L. Hennessy and David A. Patterson recently wrote a long article that concluded that the next 10 years will be a golden decade for computing architecture. The main reason is that in the past we used computers to do calculations, and now we use them to process information, but the digital universe is doubling every two years and all the energy consumption is unbearable.
Of course, there are other reasons, that is, we are now living in an era of artificial intelligence, and artificial intelligence has made great achievements. Although AlphaGo defeated the world champion of Go, artificial intelligence still has many bottlenecks. In short, the development of artificial intelligence 5 conditions must be met:
1. Sufficient data. 2. The decisive question. 3. Complete knowledge. 4. Static. 5. A single system.
For example, if you let an intelligent robot go to a destination autonomously, it cannot be done without programming it in advance. It took us a few years to establish this concept: where, how to get out, go through the door, go through the window , all of which are related to general intelligence, so our conclusion is: to develop an artificial general intelligence.
To develop artificial general intelligence, we must learn from the brain, which is currently the only general intelligence in the entire universe. Comparing the human brain with a computer, although the principles of the two systems are different, they have a strong complementary effect.
Therefore, the current computer system can be transformed by borrowing the basic principles of brain science. Developing brain-like computing is a very important part of developing artificial general intelligence, because it is the cornerstone of computing.
The development of artificial general intelligence is not a recent idea. If we look at the early articles of great scientists such as Turing and von Neumann, it is not difficult to find that this is the dream of scientists for a long time.
Why is now the best time to develop artificial general intelligence? Because with the development of sophisticated instruments, humans have learned more and more about the brain, and now it seems that we have reached a critical point in understanding the brain. The development of supercomputers can help scientists to do excellent simulations. Big data and cloud computing that save money, effort and time provide scientists with a system as complex as the brain, echoing the brain, so that we can work together. research and mutual promotion.
In addition, with the development of nanodevices, scientists can develop Electronic devices, and the energy consumption can reach the level of neurons and synapses in the human brain, so now is the best time to develop artificial general intelligence.
The development of brain-like computing to support artificial general intelligence plays a very important role in the brain. What role does it play?
Thirteen years ago, I felt that Moore’s Law would come to an end in 20 or 30 years, so I started the research on brain-like computing. Although I thought that my research was not bad, in terms of brain-like computing, I suddenly felt myself. I don’t know how to do research anymore, because there is no literature in this field, and many things need to be explored by myself, so I feel very distressed.
Once I went to climb a mountain and deliberately let myself get into the forest, no accident, I got lost. Later, I judged the direction according to the sun, stared in one direction and kept walking, walking, walking all the way to the highway, and stopped a car. Another time, I entered the forest on a cloudy day and got lost, so I thought of a way: keep climbing to the top, climb to the highest point, keep walking, keep walking, and finally reach the highway On the road, I stopped a car and went home.
Through these two things, I began to think, the brain plays the role of a compass and provides me with a sense of direction.
When doing scientific research, I like to choose fields that are more difficult to do, because I think the harder it is to do, the easier it is, because there are many competitors in the fields that are too easy, and it is difficult to lead. If it is a more difficult field, if you are doing it, no one else will do it, and you can lead, but there is a precondition: the direction must be correct. If you go on the wrong road, it will be very embarrassing.
Human intelligence is based on carbon, and we have built the current digital universe on silicon, and carbon and silicon have very similar structures, so we have a belief that what can be achieved on carbon can be achieved on silicon. Basically, it must be possible.
Discipline distribution: Challenges in developing brain-like computing and artificial general intelligence
The real challenge for developing brain-like computing and artificial general intelligence is neither science nor technology, but our disciplinary distribution. The current disciplinary distribution makes us do not have suitable people to do this research, and brain science and computer science are one Primarily exploring the natural world, the latter focuses more on applications. These two fields have different cultures, languages, and different goals, so multidisciplinary integration is especially critical.
The Brain-inspired Computing Research Center of Tsinghua University consists of 7 departments, because this field is not only the integration of computer and brain science, but also the integration of mathematics, physics, electronics, microelectronics, etc.
Teachers from our 7 departments discussed repeatedly, half a day a week, and finally we only did one thing for 7 years, called integration, integration and integration.
In this process, we have sorted out how to develop artificial general intelligence. There are two main routes: first, computer-led; second, brain science-led. Computer-led like machine learning, it has achieved brilliant results in image recognition, speech understanding, and natural language processing, but it is difficult to deal with uncertainties and so on.
Neuromorphic computing in brain science is also developing rapidly, but because we do not understand the mechanism of the brain, it has greatly hindered its development, but the two technical routes are actually complementary, and the combination of the two is currently the best in our opinion. a way.
There are actually two more to develop brain-like science: 1. Based on computers, use the basic principles of brain science to change the computing architecture; 2. We use a simple and clear word like “brain-like” to cover these two parts.
How can a brain-like computing system be built without understanding the principles of the human brain?
In this research, you actually need to study theories, chips, software, systems, cloud brains and applications. However, everyone always asks a question: If you don’t understand the human brain, how can you create a brain-like computing system?
We thought about it for a long time, and then we got the answer. The answer is this: the computer converts the information in the multi-dimensional space into a one-dimensional information flow such as 0 and 1, and uses the calculation to solve the problem. The main frequency of the CPU is getting faster and faster. In other words, you are using the time complexity. The problem is that when you reduce the dimension, the correlation is lost. This is that it is easy for people to determine whether an object is in real space or in In the mirror, but it’s hard for a computer, and that’s the root cause.
For the brain, we don’t know its basic principle, but we know that a neuron connects one thousand to ten thousand neurons. In other words, we expand the information in this place and enhance the correlation. We use the is the space complexity. In addition, the brain also uses pulses for coding, which introduces the factor of time, and we also take advantage of the complexity of space and time. Therefore, we want to maintain all the advantages of the current computer and maintain the time complexity to add a brain-like chip.
What is added? The increase is the space complexity, time and space complexity. If we look at current technology from this point of view, you will find that today’s artificial, neural network accelerators are geared towards deep artificial neural networks. It uses space complexity, and neuromorphic computing, which works like a brain, is oriented to spiking neural networks. It uses space-time complexity, a space-time complexity, a space-time complexity, why not combine it?
Therefore, we propose the Tianji chip architecture, which supports both artificial neural networks and spiking neural networks that work like brains at a cost of 3%, and also supports two heterogeneous modeling. We also use brain-like chips to build a research platform for artificial general intelligence.
Our idea is to build a multi-modal cross-research platform that can interact with the system. We use environmental changes to force the system to change. When it changes, we observe the basic principles that the system should follow to apply this change. Help us iterative development, using a Tianji chip, we can realize perception, tracking, obstacle crossing, obstacle avoidance, automatic control, voice understanding, and autonomous decision-making.
The chip is very important, and the software is also very important, because if there is no software, application engineers are reluctant to do application software development. So we developed a software tool chain ourselves, and in our lab, we have actually built the first generation of brain-like computers.
What we are doing now is a brain-like cloud brain. The difference between it and the current cloud computing is that cloud computing integrates many technologies, while the brain-like cloud brain is oriented to artificial general intelligence, because the research of artificial general intelligence is basically different from many artificial intelligence. Stacked together, our idea is to combine the elasticity of the brain with the rigidity of the computer, data-driven and knowledge-driven, and general knowledge and reasoning.
Of course, this is a very challenging long-term research, and our strategy is to proceed step-by-step. It can be imagined that we first focus on the research of one problem, which can be called the first generation, and then study the two problems together. This can be called the second generation, then the third generation, the fourth generation, and finally the fifth generation, so that we can build artificial general intelligence.
Artificial General Intelligence: Empowering All Industries
We develop brain-like computing and support artificial general intelligence. Because it is general intelligence, it can empower all walks of life and have many applications.
We are particularly interested in intelligent education, and many problems in current education can be solved through this research. For example, high-quality educational resources are scarce, resulting in inequity in education. Due to limited funds and limited equipment, it is difficult for us to truly connect theory with practice.
With the development of brain-like computing and artificial general intelligence, these will be gradually solved, and then new systems will be developed. But there is another very important factor, because education is primarily about shaping people.
Since the Industrial Revolution, human beings have developed steam engines, generators, computers, big data, and now the Internet of Everything. Humans have been changing the external world and our material life. When our material life has developed rapidly, our spiritual life has not actually developed synchronously. We are now developing brain-like computing in the age of intelligence, so that we have the opportunity to develop inward and examine our hearts.
I sincerely hope that while developing our technology and exploring the outside world, human beings can also study our inner world, cultivate both inside and outside, and develop together, so as to build a beautiful and harmonious world!
Author: Shi Luping